Alibaba’s Qwen team has released Qwen3-TTS, a text-to-speech model that the team says supports ten major languages and can adjust how it speaks based on plain-language instructions. The release extends the Qwen3 family, which already includes open-weight language, vision and audio models, into voice generation.
According to the Qwen team, Qwen3-TTS supports Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian, and ships with voice profiles reflecting several regional accents, so that developers can target applications in different markets.
Context-aware delivery
The team describes strong contextual understanding as the model’s main strength. Qwen3-TTS is designed to read the meaning of the input text and any accompanying instructions, then set tone, pace and emotional expression automatically. The team also says the model remains stable when given noisy or poorly formatted text, which matters for production use where input is rarely clean.
Users can steer the voice in natural language, describing the desired timbre, emotion or intonation directly rather than adjusting numeric parameters. The model then shapes the output to match both the instruction and the content of the text.
A 12 Hz speech tokenizer
Under the hood, the model uses a component the team calls Qwen3-TTS-Tokenizer-12Hz to represent speech. A tokenizer of this kind compresses an audio signal into a compact sequence of discrete units while preserving its meaningful features. The Qwen team says this allows a relatively lightweight architecture to reproduce high-quality speech quickly, including ambient and non-verbal sounds.
Qwen3-TTS is built as a single end-to-end system rather than a pipeline of separate stages. The team argues that this avoids the information loss and compounding errors that occur when text analysis, acoustic modeling and waveform generation are handled by separate components, and that it improves speed and overall quality across platforms.
Low-latency streaming
For real-time use, the model supports streaming generation: it can begin producing audio as soon as the first character of input arrives, and the team reports end-to-end synthesis latency of under 100 milliseconds. That figure is the team’s own measurement, but if it holds up in independent testing it would make the model suitable for live assistants and interactive voice systems where delays are noticeable.
The release continues a pattern of Chinese labs publishing capable open models across modalities. For developers, a multilingual, instruction-controllable text-to-speech model with open weights offers an alternative to proprietary voice APIs, particularly where data must stay on local infrastructure.
